[PATCH] Fixup PR78396 fix
Richard Biener
rguenther@suse.de
Thu Nov 24 12:50:00 GMT 2016
I noticed a few failing SPEC benchs on a AVX2 machine which is because
when now doing BB vectorization on a failed-to-vectorize if-converted
loop body we can end up with unvectorized masked loads/stores. As
BB vectorization does not handle masked loads/stores at all the following
patch simply avoids doing BB vectorization.
Bootstrap / regtest in progress.
I'll try to isolate a testcase later.
Richard.
2016-11-24 Richard Biener <rguenther@suse.de>
PR tree-optimization/78396
* tree-vectorizer.c (vectorize_loops): When the if-converted
body contains masked loads or stores do not attempt to
basic-block-vectorize it.
Index: gcc/tree-vectorizer.c
===================================================================
--- gcc/tree-vectorizer.c (revision 242834)
+++ gcc/tree-vectorizer.c (working copy)
@@ -570,14 +570,22 @@ vectorize_epilogue:
&& ! loop->inner)
{
basic_block bb = loop->header;
+ bool has_mask_load_store = false;
for (gimple_stmt_iterator gsi = gsi_start_bb (bb);
!gsi_end_p (gsi); gsi_next (&gsi))
{
gimple *stmt = gsi_stmt (gsi);
+ if (gimple_call_internal_p (stmt)
+ && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD
+ || gimple_call_internal_fn (stmt) == IFN_MASK_STORE))
+ {
+ has_mask_load_store = true;
+ break;
+ }
gimple_set_uid (stmt, -1);
gimple_set_visited (stmt, false);
}
- if (vect_slp_bb (bb))
+ if (! has_mask_load_store && vect_slp_bb (bb))
{
dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location,
"basic block vectorized\n");
More information about the Gcc-patches
mailing list